How do I fine tune using eager few shot object detection on custom dataset with multiple classes which is in coco format?

Question

I have been trying to use this eager few shot object detection tutorial. Instead of using the rubber ducky data, I wanted to use my custom dataset which is already in coco format and quite large. I tried to generate gt_boxes using my dataset but this tutorial only works for single class. How do I make it work for multiple classes and how do I use tf records directly to do eager few shot training in tf2.

Thanks

Answer 1

The gt_boxes usually have a shape of [N,4] where N is the number of boxes all in numpy array format. Suppose you have two classes, the gt_boxes will be [2,4] meaning you have 2 arrays all having a shape of (4,). This can be represented like so [[array([[a,b,c,d]]),array([[e,f,g,h]])]]. Running np.shape on this will yield a shape of [2,4].

You just have to figure out how do that using numpy.